Leave it as automation
The steps are the same every time, the input is structured, and a mistake is obvious. If a checklist describes the whole job, keep the checklist. An agent would only add judgement you didn't ask for.
Automation and agents
From automation to agents
I've built Zaps, Make scenarios and n8n flows for years. Now I help teams work out which ones should become agents - and which should stay boring.
People ask me to replace Zapier or n8n with AI agents. I'm an AI automation consultant, and most of the time I tell them to leave the working flows alone. An agent is for the jobs where someone is reading, judging, or drafting.
Stay or switch
What should stay automation
Plain automation is the right tool when the path doesn't change. A form lands in a sheet. A payment marks an invoice paid. A stage change pings Slack. Zapier, Make, or n8n will do that every time, and you can see when it hasn't.
An agent is the right tool when the next step depends on the content. Someone has to read the email, decide where it goes, pull a few facts together, or write a first draft. The agent can do that reading and that draft. A person still approves anything that gets sent or saved.
Worth an agent
The next step depends on what was actually said. A draft or a shortlist is useful, and a person approves it before anything is sent or saved. If getting it wrong quietly would be a problem, it doesn't run unattended.
Controls
How I keep an agent safe
Three controls ship with every agent. They're there so a person catches the mistake, not a customer.
Human approval gates
The agent can read, draft, and queue a change. It does not send the message or write back to the system until a person approves. The gate sits on the step you would not want to undo.
Error alerting
A failed run, a stall, or an answer the agent isn't sure of raises an alert. It goes to a channel someone already watches, with the input and what the agent was about to do. A log nobody opens doesn't count.
Handover runbooks
Written so someone who didn't build the agent can stop it, see what it last did, and do the job by hand. It covers where the agent lives, what it may touch, who approves, and what broken looks like.
Working together
What an engagement looks like
Audit, then one pilot, then ongoing agent ops if the pilot earns it.
Audit
I list the workflows you already run, and the manual jobs beside them. Three piles: leave it, fix it as plain automation, or try an agent. You get that split before anything is built.
First agent pilot
One workflow, kept narrow. The approval gate, the error alert, and the handover runbook are in place before it touches live work. If it misbehaves, we tighten it, or put the job back on a Zap or an n8n flow.
Ongoing agent ops
Once the pilot is behaving, I stay on it. That means adjusting the gates, fixing what breaks, and deciding whether a second agent is actually called for.
Patterns
What I've built
Three patterns of work. The client names stay off.
Staff bots
About a dozen internal staff bots for a UK luxury hotel group.
Lead enrichment
A lead-enrichment pipeline for a global design-software company. It researches and ranks decision-makers with AI before syncing to the CRM.
Questionnaire tracker
A questionnaire tracker for a US healthcare practice, built inside its compliant Google Workspace.
FAQ
Questions I get asked
Should I replace Zapier or n8n with AI agents?
Only where a person is reading, judging, or writing a first draft. A flow that copies a form into a sheet, or pings Slack when a deal stage changes, should stay a Zap or an n8n workflow. Replacing a working automation because an agent is newer gives you more ways for it to go wrong.
Which workflows should stay as plain automation?
The ones with a fixed path. If you can write every step down and none of them depend on reading the content, leave them as automation. Use an agent when the next step depends on what a message, a document, or a person actually said.
How do you keep an agent safe?
Three controls, every time. A human approval gate sits in front of anything that would send or change a record. Failures and uncertain runs raise an alert in a channel someone already watches. A handover runbook tells your team how to stop the agent and do the job by hand.
What does an engagement look like?
An audit first. I sort what you already run into leave it, fix the automation, or try an agent. Then one pilot, narrow on purpose, with the gate, the alert, and the runbook in place before it touches live work. Ongoing agent ops come after, once that pilot is behaving.
Do you work with small businesses?
Yes. I do this with small teams that already run on Zapier or n8n. If you want an agentic AI consultant for a small business, you get a decision on each workflow before anything is built.
Get in touch
Start with the audit
Tell me the workflow you're tempted to turn into an agent. I'll tell you if it should stay a Zap, and what a careful first pilot would look like.